Non-invasive stress detection using body posture and motion analysis with deep neural networks

Stress recognition often relies on physiological measurements, which provide objective and reliable markers of stress response, however, they are intrinsically invasive. This paper presents a non-invasive framework for stress recognition using vision-based body posture and motion analysis. We evaluate various 3D body posture estimation models based on deep neural networks (such as OpenPose, Channel-wise Topology Refinement Graph (GTR-Graph Convolution), extracting novel posture features that provide spatial and temporal information about the human body. Our approach integrates state-of-the-art 3D posture estimation techniques with machine learning models to develop body-based stress models. Experimental results demonstrate that the proposed method can accurately classify stress conditions, achieving a classification accuracy of up to 90.8% during social exposure and a percent of 71.5% including all experimental phases.